Troed Sångberg on Nostr: I refer to them as "thinking engines", or perhaps "reasoning engines". Just as a ...

I refer to them as "thinking engines", or perhaps "reasoning engines". Just as a human student read a lot of books with the point of being able to apply some general methodology afterwards, the same thing happens with LLMs. We managed to completely misunderstand how to use them best with that early "ask the LLM" as if they're some sort of knowledge databases.

"Modern" LLM setups pair them with vector databases so that they themselves fetch stored data exactly, they use search engines and tool repositories (MCPs) to accomplish tasks. When set up like this, "hallucinations" and slight differences between runs don't matter for their ability to complete the tasks given.

[@mmu_man](https://m.g3l.org/@mmu_man ) gunstick (npub13r5…erkj) Cassandre de l’Analyse Systémique (npub1h6x…hqvp)